Anthropic made Opus 5.5 cheaper. Then it broke four things your agent depends on.

Anthropic reduced the price of Claude Opus 5.5, lowering the cost per million input tokens from $5 to $4 and per million output tokens from $25 to $20. However, this cheaper model comes with four breaking changes that may disrupt agents relying on Opus 5. Anthropic's migration guide highlights these issues that can cause 400 errors or alter agent behavior.

Firstly, thinking controls have changed significantly. Opus 5.5 now defaults to always thinking, returning 400 errors if a request disables thinking or uses enabled with budget_tokens. Existing agents that previously disabled thinking for faster responses may need to adjust the effort level instead. Moreover, the default effort level has dropped from high to medium, potentially causing lower-cost, lower-effort operations when not explicitly set.

Secondly, forced tool calls are no longer possible. Setting tool_choice to any or tool now results in a 400 error, including on the token counting endpoint where cost estimates depend on these settings. Many agent loops force tool calls for database queries, code execution, or external services, and Anthropic's new system combined with strict tool use or structured outputs.

Thirdly, thinking blocks are now tied to the model and conversation they were produced in. On the Claude API, only Fable 5.1 and Mythos 5.1 can read Opus 5.5 thinking blocks. Routers or fallbacks that pass conversations to other models will miss the earlier reasoning instead of returning an error. Additionally, conversations must remain append-only for thinking blocks to remain valid.

An edit to earlier turns, such as trimming messages or changing tool definitions, invalidates existing blocks, and accounts created after August 31, 2026, will get errors by default.

Lastly, Opus 5.5 rejects the computer_20251124 tool and accepts computer use only through the computer_toolset_20260801 toolset. This change affects computer-use agents on both Claude API and Google Cloud, simplifying the request but requiring agents to adapt their loops to handle tool_use blocks differently. This change may go unnoticed if agents don't handle refusal reasons, resulting in tasks being cut short on the server-side.

In summary, while upgrading to Opus 5.5 offers cost savings, teams must address these breaking changes to prevent disruptions in their agents' functionality.